The choice of approximate posterior distribution plays a pivotal role in enhancing the performance of variational pedestrian detection. To enable efficient inference, most existing applications of variational inference adopt simple unimodal posterior families, such as Gaussian distributions. However, this simplification significantly limits the capability of variational detectors in complex and crowded pedestrian scenarios. To address this limitation, we introduce a novel approach based on normalizing flows called NFVPD, which extends a Gaussian posterior to a more expressive multimodal distribution. Our approximate posterior is constructed via a sequence of invertible transformations that progressively transform a simple initial density into a more complex one, allowing the number of density peak to adapt to the requirements of occluded scenes. Furthermore, we exploit the shape parameters of the normalizing flow to design a Distinguish Module, which helps determine whether a detection proposal corresponds to a duplicate prediction or a genuinely occluded pedestrian, thereby further improving detection accuracy. Extensive experiments conducted on the CrowdHuman and CityPersons datasets demonstrate the strong competitiveness of our method compared to existing state-of-the-art approaches.